Prev: Uprobes Implementation
Next: Treat as urgent and important.
From: Tim Abbott on 14 Jun 2010 23:00
On Mon, 14 Jun 2010, James Bottomley wrote:

> On Mon, 2010-06-14 at 18:21 -0400, Tim Abbott wrote:
> > On Mon, 14 Jun 2010, Matt Fleming wrote:
> >
> > > Do these special kernel sections include things like the parisc
> > > .text.do_softirq, .text.sys_exit, etc? James raised a good objection to
> > > the parisc patch of this series. I'm guessing most people saw it already
> > > but I'll paste it here for reference,
> > >
> > > This would destroy all of the named parisc text ordering we do above the
> > > removed line because now you'd have swept up all the function sections
> > > before we get to them, won't it?
> > >
> > > The ordering is an execution speed up on 32 bit systems because our
> > > relative jump is so short.
> > >
> > > Will you changes handle this OK?
> >
> > I think I addressed this in my reply to James just now, but to be super
> > clear, this -ffunction-sections plan involves renaming .text.do_softirq to
> > .text..do_softirq (etc.) first.
>
> OK, so that doesn't make a lot of sense to me; I suspect because you
> don't understand what parisc is doing. These aren't names of linux
> special sections ... they're names of function sections. For
> efficiency, we take specific hot functions and place them together in
> the linker script so the jumps between them are small enough to be coded
> as relative on the 32 bit architecture. It's really just a more
> efficient way of laying out the binary.

Yeah, I'd forgotten parisc was doing -ffunction-sections, now I understand
what you're talking about. What I'd recommend is just moving the
..text.do_softirq and friends like in the hunk below -- I believe that
should achieve the same performance goals, and should have no conflict
with adding a wildcard for -ffunction-sections generated sections to
TEXT_TEXT.

-Tim Abbott

diff --git a/arch/parisc/kernel/vmlinux.lds.S
b/arch/parisc/kernel/vmlinux.lds.S
index d64a6bb..ad0d3d3 100644
--- a/arch/parisc/kernel/vmlinux.lds.S
+++ b/arch/parisc/kernel/vmlinux.lds.S
@@ -52,15 +52,15 @@ SECTIONS
_text = .; /* Text and read-only data */
.text ALIGN(16) : {
HEAD_TEXT
+ *(.text.do_softirq)
+ *(.text.sys_exit)
+ *(.text.do_sigaltstack)
+ *(.text.do_fork)
TEXT_TEXT
SCHED_TEXT
LOCK_TEXT
KPROBES_TEXT
IRQENTRY_TEXT
- *(.text.do_softirq)
- *(.text.sys_exit)
- *(.text.do_sigaltstack)
- *(.text.do_fork)
*(.text.*)
*(.fixup)
*(.lock.text) /* out-of-line lock text */


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: James Bottomley on 16 Jun 2010 17:50
On Mon, 2010-06-14 at 22:45 -0400, Tim Abbott wrote:
> On Mon, 14 Jun 2010, James Bottomley wrote:
>
> > > I believe that the pattern [A-Za-z$_] matches all valid characters to
> > > start a function name (in particular, it includes "$"). If I'm missing
> > > any valid characters for the start of a function name, please correct me.
> >
> > Well, our linker seems to generate annoying symbols with carets in
> > them ...
>
> The question here is: is there C code that when compiled with
> -ffunction-sections will generate an ELF section with a name that starts
> with ".text.^"? For that to happen, you would need a function whose name
> started with "^", which isn't valid C.

True ... but I didn't say that it was valid C. I said the compiler is
spitting them out in the assembly ... I suspect it uses invalid C
function characters precisely to avoid clashes.

> The relevant namespace here is the names for ELF sections generated by
> -ffunction-sections. These are in turn computed by the compiler from
> function names -- there's no potential conflict created by
> linker-generated symbols whose names start with a caret. Similarly, for
> -fdata-sections, we only care about the names of C data objects, which
> also can't start with a caret.
>
> > > While one could in principle try to handle things by not renaming the
> > > .text.foo sections and instead just placing the linker script code for
> > > them all before a .text.* item in the linker script, that approach is
> > > really fragile. I think the "text..foo" approach is a good design and I
> > > am not aware of any problems with it.
> >
> > OK, but how about some actual explanation? You've just characterised
> > the current -ffunction-sections scheme that parisc has used for decades
> > as "fragile"
>
> The current parisc situation is fine.
>
> What I was trying to draw a contrast with is supporting -fdata-sections by
> adding ".data.*" to DATA_DATA, and then trying to make sure that all the
> architecture linker scripts handle all the kernel's special data sections
> with names like ".data.foo" before the place where DATA_DATA appears in
> their linker scripts. Most of the architecture linker scripts mention
> more than a half dozen special kernel sections with names of the form
> ".data.foo", often in fairly random orders, and so it would be really
> fragile to add the constraint that these sections need to all appear above
> DATA_DATA.
>
> Adding ".data.[A-Za-z$_]" to DATA_DATA doesn't have this problem.
>
> If we similarly added ".text.[A-Za-z$_]" to TEXT_TEXT, we'd presumably
> move the 4 named .text.foo sections before TEXT_TEXT; I don't think any
> other architectures would require any work.
>
> > > Some more detailed explanation is available here:
> > > <http://lkml.org/lkml/2010/2/19/365>
> >
> > That's still a bit short on explanations.
> >
> > But if I infer from the rest, someone, somewhere broke the convention
> > that all our special linux sections be called .XX.data and .XX.text to
> > distinguish them from the .text.FF and .data.YY the compiler will
> > generate with the relevant sectional directives? because it's been
> > working OK for us for a while.
>
> I don't know the full history here. But prior to the ".data.foo" ->
> ".data..foo" patches that were merged recently, there were a bunch of
> cross-architecture sections with these sorts of names, e.g.:
>
> .data.page_aligned
> .data.nosave
> .data.read_mostly
> .data.cacheline_aligned
> .data.lock_aligned
> .data.percpu*
> .data.init_task
> etc.
>
> There were also a bunch of ".text.foo" sections on individual
> architectures, many of which currently don't support -ffunction-sections
> (sh, ia64, x86, mips, etc.).
>
> However, there weren't any .text.foo sections that are cross-architecture.
> Since parisc only uses -ffunction-sections, and not -fdata-sections, the
> popular .data.foo naming scheme doesn't cause any breakage on parisc.
>
> The only architecture that does use -fdata-sections is frv, and there
> could theoretically have been breakage there, but in practice it's likely
> nobody has written kernel code that would actually conflict, e.g. "static
> int percpu = 3;", yet.

Right, but what I was curious about, since, we already sorted all the
problems with text sections out on parisc by putting the .text as a
suffix not a prefix, and you have to change all the data sections
anyway, why not just follow established practise?

> > To fix the breakage, the proposal now is to name all linux special
> > sections as .text..XX and .data..XX? I can see that's more standard
> > looking that XX.text and XX.data, but not necessarily that it's better.
>
> Yes, that's the proposal.
>
> > This then introduces a problem of matching because .text.X and .text..X
> > are hard to distinguish using the linker matching scripts.
>
> Right. I believe that this is totally solvable with a simple linker
> script pattern, since the space of valid names for functions and data
> objects in C code is quite restricted (and that the implementation of
> using e.g. ".data.[A-Za-z$_]*" solves this problem).
>
> > So even if I buy the rename of the linux symbols, what about using a
> > linker defined symbol that's illegal as a function as the initial
> > separator instead of .? So hyphen looks the obvious one ... you can
> > have all the linux special sections being .text- and .data- then we can
> > easily distinguish.
>
> Is "." a valid first character for a function name? I don't see the
> problem with using "." here.

The problem is that it's hard to get the linker to distinguish
between .text.. and .text. without funny regexp contortions. However,
if the two sections began .text- and .text. they are easy to distinguish
because one prefix isn't a subset of the other.

> Both .page_aligned.data and .data-page_aligned are valid choices (and in
> fact, the first patch series I sent on this topic about 18 months ago did
> .page_aligned.data, I think).
>
> The main technical difference between ".data..page_aligned" and
> ".page_aligned.data" in my view is that you need to be more careful when
> writing assembly files with ".page_aligned.data".
>
> To give an example, if I run the following:
>
> $ cat > foo.s
> .section .data-page-aligned
> .long 0
> .section .data.page_aligned
> .long 1
> $ gcc -c foo.s -o foo.o
> $ objdump -h foo.o
> foo.o: file format elf32-i386
>
> Sections:
> Idx Name Size VMA LMA File off Algn
> 0 .text 00000000 00000000 00000000 00000034 2**2
> CONTENTS, ALLOC, LOAD, READONLY, CODE
> 1 .data 00000000 00000000 00000000 00000034 2**2
> CONTENTS, ALLOC, LOAD, DATA
> 2 .bss 00000000 00000000 00000000 00000034 2**2
> ALLOC
> 3 .data-page-aligned 00000004 00000000 00000000 00000034 2**0
> CONTENTS, READONLY
> 4 .data.page_aligned 00000004 00000000 00000000 00000038 2**0
> CONTENTS, ALLOC, LOAD, DATA
>
> one can see that the .data-page-aligned section doesn't have the right
> section flags. So I'm pretty sure the relevant assembler heuristic is
> looking for things starting with ".data.", not just ".data".
>
> The kernel has a lot of code in assembly files that just does:
>
> .section ".data"
>
> and so there's a very real risk that folks who are doing pattern-matching
> development on assembly files will end up creating non-allocated sections
> by accident (I've certainly made this mistake myself, and there's a bug of
> this form in arch/x86/lib/thunk.S until commit
> c6c2d7a084d14a8a701be84872aa1b77d2945f46, so I don't think I'm alone)

But that commit isn't anything to do with the section not being
allocated. That's actually irrelevant to us, since we sort out the
sectional allocations out at link time with the linker script (which
ends up completely ignoring the original file flags). The problem,
specifically, is that the linker does a rename of two identically named
sections with different flags ... we would also get the same behaviour
if the linker thought two sections weren't mergeable even if they did
have identical flags.

In general, I think it's good practise for all use of sections to
specify the elf flags.

> I also think that ".data..page_aligned" is more readable as a new name for
> the former ".data.page_aligned" than ".page_aligned.data" is, but I think
> that's a secondary consideration. ".data.-page_aligned" would be
> technically equivalent to ".data..page_aligned", but I think it is uglier.

Actually, as I said, that would be .data-

But, to be honest, I don't care that much about the data names because
we don't use -fdata-sections, the only objection is that having two
separate conventions for text and data is adding needless complexity ...
I do care about the text names, and I'd rather we stuck to the existing
convention there, or provided a good reason for the change ... and
wanting .text. as a prefix for everything is weak at best.

The specific objection I have to converting everything to the .text..*
prefix for linux sections is that the gather all function sections can
no longer be *(.text.*), but has to become a regular expression, which
is adding fragility.

James


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Denys Vlasenko on 17 Jun 2010 15:00
On Tue, Jun 15, 2010 at 12:02 AM, Tim Abbott <tabbott(a)ksplice.com> wrote:
> On Mon, 14 Jun 2010, James Bottomley wrote:
>
>> On Mon, 2010-06-14 at 20:33 +0100, Matt Fleming wrote:
>> > On Mon, 14 Jun 2010 10:32:46 -0400 (EDT), Tim Abbott <tabbott(a)ksplice.com> wrote:
>> > >
>> > > I was planning to submit in the next couple weeks a change that adds
>> > > support for building the kernel with -ffunction-sections -fdata-sections,
>> > > which would have as a piece of it adding to TEXT_TEXT the following
>> > > expression:
>> > >
>> > > � *(.text.[A-Za-z$_]*) � �/* handle -ffunction-sections */\
>>
>> Just as a point of technical interest, that won't handle
>> -ffunction-sections. �At least on parisc, we get a
>> section .text.<function name> for every function. �This means that any
>> character legal in a function name can appear there, not just letters
>> and underscores (we get millicode ones with dollar signs as well for
>> instance). �That's why *(.text.*) is safer
>
> Hi James,
>
> I believe that the pattern [A-Za-z$_] matches all valid characters to
> start a function name (in particular, it includes "$"). �If I'm missing
> any valid characters for the start of a function name, please correct me.

Yes, I think we need to add 0-9 too. C names can't have a digit as a starting
character, but linker can produce such names when invoked as "ld -r --unique"
(incremental linking).

Currently we use just "ld -r" to combine all .o files from a directory
into one bigger .o file, but this combines all similarly-named sections.
This not only combines all .text sections from every input .o file
into one .text section (which isn't surprising), but also combines
all .text.func sections too. Which we don't want to happen when we
(eventually) want to link kernel with --gc-sections.

The fix already exists: "ld -r --unique". With --unique, ld
will create unique sections named .text.1, .text.2, .text.func.1
and such.

Therefore, in order to accomodate .text.NUM sections in the future,
we'd better use .text.[A-Za-z0-9$_]* pattern.

--
vda
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: Denys Vlasenko on 17 Jun 2010 15:20
On Wed, Jun 16, 2010 at 11:40 PM, James Bottomley
<James.Bottomley(a)hansenpartnership.com> wrote:
> True ... but I didn't say that it was valid C. �I said the compiler is
> spitting them out in the assembly ... I suspect it uses invalid C
> function characters precisely to avoid clashes.

Then we need to add all such chars to that regexp.

> The problem is that it's hard to get the linker to distinguish
> between .text.. and .text. without funny regexp contortions. �However,
> if the two sections began .text- and .text. they are easy to distinguish
> because one prefix isn't a subset of the other.
>
>> Both .page_aligned.data and .data-page_aligned are valid choices (and in
>> fact, the first patch series I sent on this topic about 18 months ago did
>> .page_aligned.data, I think).
>>
>> The main technical difference between ".data..page_aligned" and
>> ".page_aligned.data" in my view is that you need to be more careful when
>> writing assembly files with ".page_aligned.data".
>>
>> To give an example, if I run the following:
>>
>> $ cat > foo.s
>> .section .data-page-aligned
>> � � � .long 0
>> .section .data.page_aligned
>> � � � .long 1
>> $ gcc -c foo.s -o foo.o
>> $ objdump -h foo.o
>> foo.o: � � file format elf32-i386
>>
>> Sections:
>> Idx Name � � � � �Size � � �VMA � � � LMA � � � File off �Algn
>> � 0 .text � � � � 00000000 �00000000 �00000000 �00000034 �2**2
>> � � � � � � � � � CONTENTS, ALLOC, LOAD, READONLY, CODE
>> � 1 .data � � � � 00000000 �00000000 �00000000 �00000034 �2**2
>> � � � � � � � � � CONTENTS, ALLOC, LOAD, DATA
>> � 2 .bss � � � � �00000000 �00000000 �00000000 �00000034 �2**2
>> � � � � � � � � � ALLOC
>> � 3 .data-page-aligned 00000004 �00000000 �00000000 �00000034 �2**0
>> � � � � � � � � � CONTENTS, READONLY
>> � 4 .data.page_aligned 00000004 �00000000 �00000000 �00000038 �2**0
>> � � � � � � � � � CONTENTS, ALLOC, LOAD, DATA
>>
>> one can see that the .data-page-aligned section doesn't have the right
>> section flags. �So I'm pretty sure the relevant assembler heuristic is
>> looking for things starting with ".data.", not just ".data".
>>
>> The kernel has a lot of code in assembly files that just does:
>>
>> .section ".data"
>>
>> and so there's a very real risk that folks who are doing pattern-matching
>> development on assembly files will end up creating non-allocated sections
>> by accident (I've certainly made this mistake myself, and there's a bug of
>> this form in arch/x86/lib/thunk.S until commit
>> c6c2d7a084d14a8a701be84872aa1b77d2945f46, so I don't think I'm alone)
>
> But that commit isn't anything to do with the section not being
> allocated. �That's actually irrelevant to us, since we sort out the
> sectional allocations out at link time with the linker script (which
> ends up completely ignoring the original file flags). �The problem,
> specifically, is that the linker does a rename of two identically named
> sections with different flags ... we would also get the same behaviour
> if the linker thought two sections weren't mergeable even if they did
> have identical flags.
>
> In general, I think it's good practise for all use of sections to
> specify the elf flags.

This is doable in asm, yes. For .bss, we need to not forget about
@nobits too: section .bss.foo,"aw",@nobits

> Actually, as I said, that would be .data-
>
> But, to be honest, I don't care that much about the data names because
> we don't use -fdata-sections, the only objection is that having two
> separate conventions for text and data is adding needless complexity ...
> I do care about the text names, and I'd rather we stuck to the existing
> convention there, or provided a good reason for the change ... and
> wanting .text. as a prefix for everything is weak at best.
>
> The specific objection I have to converting everything to the .text..*
> prefix for linux sections is that the gather all function sections can
> no longer be *(.text.*), but has to become a regular expression, which
> is adding fragility.

I agree that it's not pretty, but otherwise we add fragility in other places:
every section directive must be correct now. These places are more numerous
than linker scripts.

But, leaving aside the question whether we want to take the risk
of people forgetting to do it properly everywhere, -
is it even *possible* to specify these attributes in C - that is,
in gcc's __attribute__((section(".bss.foo")))?

I think currently gcc does not support it - it guesses attributes
based on the section name.

--
vda
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: James Bottomley on 17 Jun 2010 16:00
On Thu, 2010-06-17 at 21:11 +0200, Denys Vlasenko wrote:
> On Wed, Jun 16, 2010 at 11:40 PM, James Bottomley
> <James.Bottomley(a)hansenpartnership.com> wrote:
> > True ... but I didn't say that it was valid C. I said the compiler is
> > spitting them out in the assembly ... I suspect it uses invalid C
> > function characters precisely to avoid clashes.
>
> Then we need to add all such chars to that regexp.

Thus proving the point about fragility ...

> > The problem is that it's hard to get the linker to distinguish
> > between .text.. and .text. without funny regexp contortions. However,
> > if the two sections began .text- and .text. they are easy to distinguish
> > because one prefix isn't a subset of the other.
> >
> >> Both .page_aligned.data and .data-page_aligned are valid choices (and in
> >> fact, the first patch series I sent on this topic about 18 months ago did
> >> .page_aligned.data, I think).
> >>
> >> The main technical difference between ".data..page_aligned" and
> >> ".page_aligned.data" in my view is that you need to be more careful when
> >> writing assembly files with ".page_aligned.data".
> >>
> >> To give an example, if I run the following:
> >>
> >> $ cat > foo.s
> >> .section .data-page-aligned
> >> .long 0
> >> .section .data.page_aligned
> >> .long 1
> >> $ gcc -c foo.s -o foo.o
> >> $ objdump -h foo.o
> >> foo.o: file format elf32-i386
> >>
> >> Sections:
> >> Idx Name Size VMA LMA File off Algn
> >> 0 .text 00000000 00000000 00000000 00000034 2**2
> >> CONTENTS, ALLOC, LOAD, READONLY, CODE
> >> 1 .data 00000000 00000000 00000000 00000034 2**2
> >> CONTENTS, ALLOC, LOAD, DATA
> >> 2 .bss 00000000 00000000 00000000 00000034 2**2
> >> ALLOC
> >> 3 .data-page-aligned 00000004 00000000 00000000 00000034 2**0
> >> CONTENTS, READONLY
> >> 4 .data.page_aligned 00000004 00000000 00000000 00000038 2**0
> >> CONTENTS, ALLOC, LOAD, DATA
> >>
> >> one can see that the .data-page-aligned section doesn't have the right
> >> section flags. So I'm pretty sure the relevant assembler heuristic is
> >> looking for things starting with ".data.", not just ".data".
> >>
> >> The kernel has a lot of code in assembly files that just does:
> >>
> >> .section ".data"
> >>
> >> and so there's a very real risk that folks who are doing pattern-matching
> >> development on assembly files will end up creating non-allocated sections
> >> by accident (I've certainly made this mistake myself, and there's a bug of
> >> this form in arch/x86/lib/thunk.S until commit
> >> c6c2d7a084d14a8a701be84872aa1b77d2945f46, so I don't think I'm alone)
> >
> > But that commit isn't anything to do with the section not being
> > allocated. That's actually irrelevant to us, since we sort out the
> > sectional allocations out at link time with the linker script (which
> > ends up completely ignoring the original file flags). The problem,
> > specifically, is that the linker does a rename of two identically named
> > sections with different flags ... we would also get the same behaviour
> > if the linker thought two sections weren't mergeable even if they did
> > have identical flags.
> >
> > In general, I think it's good practise for all use of sections to
> > specify the elf flags.
>
> This is doable in asm, yes. For .bss, we need to not forget about
> @nobits too: section .bss.foo,"aw",@nobits

That's only for bss ... we have about a handful of such statements and
they always use the assembler .bss directive (which doesn't need flags).

> > Actually, as I said, that would be .data-
> >
> > But, to be honest, I don't care that much about the data names because
> > we don't use -fdata-sections, the only objection is that having two
> > separate conventions for text and data is adding needless complexity ...
> > I do care about the text names, and I'd rather we stuck to the existing
> > convention there, or provided a good reason for the change ... and
> > wanting .text. as a prefix for everything is weak at best.
> >
> > The specific objection I have to converting everything to the .text..*
> > prefix for linux sections is that the gather all function sections can
> > no longer be *(.text.*), but has to become a regular expression, which
> > is adding fragility.
>
> I agree that it's not pretty, but otherwise we add fragility in other places:
> every section directive must be correct now. These places are more numerous
> than linker scripts.

I thought I just refuted that in the above: we don't care what the
assembler sections are flagged as because the linker script gets to pick
the flags anyway ... so most bugs arrived at this way have no visible
side effects ... and section merging problems have to be accounted for
anyway in the final linker scripts.

James

> But, leaving aside the question whether we want to take the risk
> of people forgetting to do it properly everywhere, -
> is it even *possible* to specify these attributes in C - that is,
> in gcc's __attribute__((section(".bss.foo")))?
>
> I think currently gcc does not support it - it guesses attributes
> based on the section name.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
First  |  Prev  |  Next  |  Last
Pages: 1 2 3 4
Prev: Uprobes Implementation
Next: Treat as urgent and important.